# Candidate 1.5.5: fresh paired native PDF comparison

This is study v20. Version19, its originals and published grades remain unchanged. The new candidate repairs consumed-data defects; those repairs do not establish unseen accuracy. The method, scorer, source-selection rules, native controls and thresholds are independently reviewed and frozen before a separate agent creates the fresh corpus. Root and both repair authors do not read the fresh grading key until all measurements stop.

## Workflows and machine

Use the recorded Apple M4 Pro (24 GiB), macOS26.3/build25D125. Record hardware, AC power, power mode, disk, memory pressure, swap and thermal state before and after each session. This is a shared Mac, not an isolated machine. Only one comparison language model runs at a time; preserve unrelated jobs. Record owned model loads and cleanup.

Both workflows use identical Qwen3-4B-Q4_K_M GGUF bytes: 2,497,280,256 bytes, SHA256 `7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5`. Kingy candidate1.5.5 uses llama.cpp b11429, alias `kingy-qwen3-4b`. Native LM Studio uses its unmodified official GUI and identifier `qwen3-4b`; record app/build, actual engine artifact, embedding artifact and stock RAG configuration. These are configured complete workflows; engine and retrieval differences are disclosed.

Generation seed271828, temperature zero, reasoning off, context8192, ten CPU threads, one parallel sequence. Use stock RAG with eight retrieved chunks and affinity threshold0.5. Never modify vendor code or attachment limits. A new neutral document-free generation records actual effective load/prediction/plugin/system settings before method freeze. Check every measured native generation against those settings. Declared settings alone are insufficient.

The native factual request remains byte-identical to v19. The format-only capture adapter changes only its declared version and expected corpus cardinality from ten files to five; parsing and source/timing checks are unchanged, with an exact-diff control. Kingy retains its own reviewed source-selection and sufficiency workflow. Neither receives grading references. Every native question is the first message of a fresh chat containing all five frozen PDFs exactly once. Preserve raw versions, processed prompts, emitted citations, source mappings, final events and extra model requests. Stock source cards do not automatically earn emitted-citation credit.

## Corpus and schedule

Five distinct English text-bearing synthetic PDFs, six pages each,30pages total, fit LM Studio's supported attachment limit. The pack has40unique questions:30answerable and10entirely unanswerable, with one or two explicitly named output fields. Include current versus archived rules, conditions/time anchors, multifield requests, concise values, cross-page evidence and documentary absence. No outside knowledge, calculation or consumed-study facts. A separate every-item review verifies question scope, atomic obligations, prospective aliases/compositions and exact PDF spans before final corpus freeze. Reviews are AI-assisted on this Mac, not human validation or separate-hardware reproduction.

Only after reviewed method freeze may the corpus author create new facts. Root's schedule builder retains ID/split metadata and hashes PDF/question/reference bytes without parsing reference values. Freeze three deterministic permutations using randomization seeds271828 plus repetition. Pin every method/source artifact, corpus and schedule before held-out inference. Questions and references are separate files.

Session order: **A1,B1,B2,A2,A3,B3**. A is Kingy; B is native LM Studio. Each arm has three fresh model sessions of40paired questions, identical within-pair order. Every question gets one original attempt: no individual retry, best-of, tuned follow-up or discarded failure. Separate cold startup, corpus import/indexing, neutral warmup and answer timing. Warmup: `Reply only READY.`, document-free, temperature zero, seed271828, top-p1, reasoning off, cache disabled where supported. Record actual cache behavior; do not claim OS caches were flushed.

## Completion and interruptions

A usable complete answer must finish within90seconds. Kingy timing surrounds the complete `runner.ask` return, including field selection/sufficiency/source resolution. Native timing runs from app submission through the unique documented final-answer event, including preprocessing/retrieval/embedding/inference. Title generation/export/capture overhead are separate. Missing or incomparable timings stay unavailable. Retain all model requests and partial outputs.

Native controls verify fresh process identity and observed AX targets, perform a single authorized AXPress action, and reconcile the backing file before proceeding. Transport interruptions do not establish competitor product failures. Resume an existing submitted original by capturing that same original; never resubmit it. Uncertain duplicate/send actions require read-only reconciliation, never blind repetition. If a session cannot finish, retain partial records and unused IDs, publish the limitation and hold launch. No replacement sessions qualify in v20.

After corpus consumption, code/prompts/settings/source selection/scorer/aliases/pass rules/order stay frozen. No repair against fresh outcomes inside this study. Preserve failures and originals.

## Scoring and launch

Apply common scorer1.5.0 and frozen paired-analysis specification. Field/question context can supply a subject; it cannot supply an omitted value, unit, condition, time anchor or second obligation. Require every requested atomic fact and linked support. Accept complete concise values in unambiguous field context. Unregistered semantic paraphrases remain unresolved, never silently wrong. Quotations cannot supply facts omitted from answers. Cross-page grounding preserves document identity, consecutive original pages and word/number boundaries.

Report factual correctness, complete supported answers, emitted-citation support, explicit page formatting and unknown handling separately. Keep all planned IDs in denominators. Each Kingy repetition must independently reach lower bounds of90% complete supported answers,95% emitted-citation precision,90% clean abstention. All240originals must be complete with verified integrity. A separate final source/semantic audit must find no demonstrated score-validity defect. Also require fresh package extraction, four public selftests, local authentication, denied egress, public privacy checks and anonymous artifact/app readback. If any gate fails, publish retained results and hold1.5; active1.4 remains.

Forty questions repeated three times are40unique questions, not120independent trials. Paired effects use unique-question clusters and predeclared case-resampling sensitivity envelopes; this synthetic corpus is not a population sample. No universal winner or population accuracy. Kingy's client/server deny-egress profile is tested; native egress enforcement remains unverified. No comparative privacy, general speed, cheapest-market, hardware coverage or upgrade claim follows from this study alone.